OpenAI Rolls Out Text Watermarking for EU Under AI Act

OpenAI is deploying text watermarking for ChatGPT and Codex outputs in the EU over the coming weeks to comply with Article 50 of the EU AI Act, which requires machine-readable marks on generative AI outputs. The watermark, called textGrain, works by statistically tilting the model's word choices to embed a hidden signal. The rollout extends existing provenance tools that already cover images and audio. API customers globally can opt in to text watermarking starting October 5, 2026. OpenAI acknowledges significant limitations: at a 1% false positive rate, the detector caught roughly 80% of 200-token passages and 95% of 400-token ones, with math-heavy text scoring substantially lower. Editing erodes detection substantially, with replacing 25% of words with synonyms in 400-token passages dropping detection from around 92% to 17%. The public cannot access the watermark detector; it is restricted to OpenAI and an approved group of researchers and expert organizations.